Automated Contusion Volume Measurements from Acute Phase CT Predict Post-traumatic Epilepsy
Bilal Ayvaz1, Justin Wheelock1, Daniel Jin1, Jenna Appleton2, M. Arsalan Bashir1, Khai Theeng Chow1, Tyler Frantz1, Samuel Snider3, Lawrence Hirsch1, Adithya Sivaraju1, Brandon Westover4, Sahar F. Zafar5, Aaron F. Struck6, Sacit Bulent Omay2, Brian Edlow5, Emily Gilmore1, Jennifer Kim1
1Department of Neurology, 2Department of Neurosurgery, Yale School of Medicine, 3Massachusetts General Hospital, Brigham, Harvard, 4Department of Neurology, Beth Israel Deaconess Medical Center, 5Department of Neurology, Massachusetts General Hospital, 6Department of Neurology, University of Wisconsin Hospital and Clinics
Objective:
Evaluate whether convolutional neural network (CNN)-based automated quantitative contusion measurements can predict the development of post-traumatic epilepsy (PTE).
Background:
PTE is a long-term complication of traumatic brain injury (TBI). CT is the most commonly used neuroimaging modality following acute TBI, and prior studies have linked larger contusions on CT, such as hemorrhage and edema, to a higher risk of PTE. However, manual segmentation of these contusions is time-consuming and requires skilled professionals. By leveraging a CNN-based algorithm, we aim to evaluate the predictive value of automated contusion calculations for PTE prediction.
Design/Methods:
We retrospectively identified and matched 50 PTE patients with 50 non-PTE controls based on TBI severity (GCS at admission), age, and sex. We obtained clinical variables and CT images from patients at admission and at the time hemorrhagic stability was first noted. Utilizing the CNN-based algorithm Blast-CT (2020), we automatically measured volumes for intraparenchymal hemorrhage (IPH), edema, intraventricular hemorrhage (IVH), and extra-axial hemorrhage (EAH) from admission and stability CTs. We performed univariate logistic regressions on clinical and neuroimaging variables; all univariable features p<0.1 were incorporated into a multivariate logistic regression model. Finally, we used leave-one-out cross-validation to evaluate our model’s performance.
Results:
Our univariate analysis showed that early seizures (p=0.01), neurosurgical intervention within the first 7 days of injury (p<0.01), IPH volume on admission (p=0.02) and edema volume on admission (p<0.01) were associated with PTE. Univariable analyses of images at time of hemorrhage stability also showed IPH volume (p<0.01), edema (p<0.01), and combined IPH+edema (p<0.01) were significantly associated with PTE. Our multivariate model resulted in an AUC of 0.73 (95%CI:0.63–0.83).
Conclusions:
Our results show that automatically measured contusion volumes from acute CT images can predict PTE. Future studies with larger cohorts and comparison with other algorithms are needed to valid our model.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.